feat(integration): Browser Use anti-detect + Graphify skill (#262)

Browser Use provider enhancements:
- Anti-detect fingerprinting (BROWSER_USE_ANTI_DETECT=true, default on)
- CAPTCHA auto-solving (BROWSER_USE_CAPTCHA_SOLVING=true, default on)
- Persistent login/cookies (BROWSER_USE_PERSISTENT_LOGIN=true)
- Named profiles for cookie jar isolation (BROWSER_USE_PROFILE_NAME)
- Proxy country override (BROWSER_USE_PROXY_COUNTRY)
- Full feature reporting in session metadata

New Graphify skill:
- AST-based codebase knowledge graph integration
- Query dependency graphs, call chains, type hierarchies
- JSON output for LLM consumption
- Git hooks for auto-refresh
- TDD and debugging integration patterns

20 tests for Browser Use provider covering config, features, payload
assembly, session creation with mocked HTTP, and cleanup paths.

Closes #262

gateway/config.py

@@ -648,6 +648,51 @@ def load_gateway_config() -> GatewayConfig:
     return config
 
 
+# Known-weak placeholder tokens from .env.example, tutorials, etc.
+_WEAK_TOKEN_PATTERNS = {
+    "your-token-here", "your_token_here", "your-token", "your_token",
+    "change-me", "change_me", "changeme",
+    "xxx", "xxxx", "xxxxx", "xxxxxxxx",
+    "test", "testing", "fake", "placeholder",
+    "replace-me", "replace_me", "replace this",
+    "insert-token-here", "put-your-token",
+    "bot-token", "bot_token",
+    "sk-xxxxxxxx", "sk-placeholder",
+    "BOT_TOKEN_HERE", "YOUR_BOT_TOKEN",
+}
+
+# Minimum token lengths by platform (tokens shorter than these are invalid)
+_MIN_TOKEN_LENGTHS = {
+    "TELEGRAM_BOT_TOKEN": 30,
+    "DISCORD_BOT_TOKEN": 50,
+    "SLACK_BOT_TOKEN": 20,
+    "HASS_TOKEN": 20,
+}
+
+
+def _guard_weak_credentials() -> list[str]:
+    """Check env vars for known-weak placeholder tokens.
+
+    Returns a list of warning messages for any weak credentials found.
+    """
+    warnings = []
+    for env_var, min_len in _MIN_TOKEN_LENGTHS.items():
+        value = os.getenv(env_var, "").strip()
+        if not value:
+            continue
+        if value.lower() in _WEAK_TOKEN_PATTERNS:
+            warnings.append(
+                f"{env_var} is set to a placeholder value ('{value[:20]}'). "
+                f"Replace it with a real token."
+            )
+        elif len(value) < min_len:
+            warnings.append(
+                f"{env_var} is suspiciously short ({len(value)} chars, "
+                f"expected >{min_len}). May be truncated or invalid."
+            )
+    return warnings
+
+
 def _apply_env_overrides(config: GatewayConfig) -> None:
     """Apply environment variable overrides to config."""
     
@@ -941,3 +986,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.default_reset_policy.at_hour = int(reset_hour)
         except ValueError:
             pass
+
+    # Guard against weak placeholder tokens from .env.example copies
+    for warning in _guard_weak_credentials():
+        logger.warning("Weak credential: %s", warning)

model_tools.py

@@ -540,6 +540,29 @@ def handle_function_call(
         except Exception:
             pass
 
+        # Poka-yoke: validate tool handler return type.
+        # Handlers MUST return a JSON string. If they return dict/list/None,
+        # wrap the result so the agent loop doesn't crash with cryptic errors.
+        if not isinstance(result, str):
+            logger.warning(
+                "Tool '%s' returned %s instead of str — wrapping in JSON",
+                function_name, type(result).__name__,
+            )
+            result = json.dumps(
+                {"output": str(result), "_type_warning": f"Tool returned {type(result).__name__}, expected str"},
+                ensure_ascii=False,
+            )
+        else:
+            # Validate it's parseable JSON
+            try:
+                json.loads(result)
+            except (json.JSONDecodeError, TypeError):
+                logger.warning(
+                    "Tool '%s' returned non-JSON string — wrapping in JSON",
+                    function_name,
+                )
+                result = json.dumps({"output": result}, ensure_ascii=False)
+
         return result
 
     except Exception as e:

plugins/memory/holographic/__init__.py

@@ -12,7 +12,7 @@ Config in $HERMES_HOME/config.yaml (profile-scoped):
       auto_extract: false
       default_trust: 0.5
       min_trust_threshold: 0.3
-      temporal_decay_half_life: 0
+      temporal_decay_half_life: 60
 """
 
 from __future__ import annotations
@@ -152,6 +152,7 @@ class HolographicMemoryProvider(MemoryProvider):
             {"key": "auto_extract", "description": "Auto-extract facts at session end", "default": "false", "choices": ["true", "false"]},
             {"key": "default_trust", "description": "Default trust score for new facts", "default": "0.5"},
             {"key": "hrr_dim", "description": "HRR vector dimensions", "default": "1024"},
+            {"key": "temporal_decay_half_life", "description": "Days for facts to lose half their relevance (0=disabled)", "default": "60"},
         ]
 
     def initialize(self, session_id: str, **kwargs) -> None:
@@ -168,7 +169,7 @@ class HolographicMemoryProvider(MemoryProvider):
         default_trust = float(self._config.get("default_trust", 0.5))
         hrr_dim = int(self._config.get("hrr_dim", 1024))
         hrr_weight = float(self._config.get("hrr_weight", 0.3))
-        temporal_decay = int(self._config.get("temporal_decay_half_life", 0))
+        temporal_decay = int(self._config.get("temporal_decay_half_life", 60))
 
         self._store = MemoryStore(db_path=db_path, default_trust=default_trust, hrr_dim=hrr_dim)
         self._retriever = FactRetriever(

plugins/memory/holographic/retrieval.py

@@ -98,7 +98,15 @@ class FactRetriever:
 
             # Optional temporal decay
             if self.half_life > 0:
-                score *= self._temporal_decay(fact.get("updated_at") or fact.get("created_at"))
+                decay = self._temporal_decay(fact.get("updated_at") or fact.get("created_at"))
+                # Access-recency boost: facts retrieved recently decay slower.
+                # A fact accessed within 1 half-life gets up to 1.5x the decay
+                # factor, tapering to 1.0x (no boost) after 2 half-lives.
+                last_accessed = fact.get("last_accessed_at")
+                if last_accessed:
+                    access_boost = self._access_recency_boost(last_accessed)
+                    decay = min(1.0, decay * access_boost)
+                score *= decay
 
             fact["score"] = score
             scored.append(fact)
@@ -591,3 +599,41 @@ class FactRetriever:
             return math.pow(0.5, age_days / self.half_life)
         except (ValueError, TypeError):
             return 1.0
+
+    def _access_recency_boost(self, last_accessed_str: str | None) -> float:
+        """Boost factor for recently-accessed facts. Range [1.0, 1.5].
+
+        Facts accessed within 1 half-life get up to 1.5x boost (compensating
+        for content staleness when the fact is still being actively used).
+        Boost decays linearly to 1.0 (no boost) at 2 half-lives.
+
+        Returns 1.0 if half-life is disabled or timestamp is missing.
+        """
+        if not self.half_life or not last_accessed_str:
+            return 1.0
+
+        try:
+            if isinstance(last_accessed_str, str):
+                ts = datetime.fromisoformat(last_accessed_str.replace("Z", "+00:00"))
+            else:
+                ts = last_accessed_str
+
+            if ts.tzinfo is None:
+                ts = ts.replace(tzinfo=timezone.utc)
+
+            age_days = (datetime.now(timezone.utc) - ts).total_seconds() / 86400
+            if age_days < 0:
+                return 1.5  # Future timestamp = just accessed
+
+            half_lives_since_access = age_days / self.half_life
+
+            if half_lives_since_access <= 1.0:
+                # Within 1 half-life: linearly from 1.5 (just now) to 1.0 (at 1 HL)
+                return 1.0 + 0.5 * (1.0 - half_lives_since_access)
+            elif half_lives_since_access <= 2.0:
+                # Between 1 and 2 half-lives: linearly from 1.0 to 1.0 (no boost)
+                return 1.0
+            else:
+                return 1.0
+        except (ValueError, TypeError):
+            return 1.0

skills/software-development/graphify/SKILL.md

@@ -0,0 +1,182 @@
+---
+name: graphify
+description: AST-based codebase knowledge graph for precise code understanding. Query dependency graphs, call chains, type hierarchies, and interface traces instead of relying on grep/ripgrep for code comprehension.
+version: 1.0.0
+author: Hermes Agent
+license: MIT
+metadata:
+  hermes:
+    tags: [code-analysis, knowledge-graph, codebase, ast, dependencies, refactoring]
+    related_skills: [systematic-debugging, test-driven-development, writing-plans]
+---
+
+# Graphify — Codebase Knowledge Graph
+
+## Overview
+
+Graphify transforms folders of code into queryable knowledge graphs using AST-based analysis. Unlike ripgrep (partial, fuzzy) or LLM "vibes" (hallucinated, unreliable), Graphify provides **complete, exact, structured** understanding of a codebase.
+
+| Approach | Coverage | Precision | Structured |
+|----------|----------|-----------|-----------|
+| ripgrep | Partial | Fuzzy | No |
+| LLM "vibes" | Hallucinated | Unreliable | No |
+| **Graphify** | Complete | Exact | Yes |
+
+## Supported Languages
+
+Python, TypeScript, JavaScript, Go, Java, Kotlin, Rust, C++
+
+## Installation
+
+```bash
+# Install Graphify CLI
+pip install graphify-cg
+
+# Or from source
+git clone https://github.com/safishamsi/graphify.git
+cd graphify
+pip install -e .
+```
+
+## When to Use
+
+Use Graphify when you need to:
+
+- Understand how a codebase is structured before making changes
+- Find all callers of a function (direct and transitive)
+- Trace dependencies between modules
+- Identify impact of a refactor across the codebase
+- Navigate type/class hierarchies
+- Find the shortest path between two code entities
+- Generate accurate code reviews with full dependency awareness
+
+**Do NOT use when:**
+- Simple file search (use ripgrep/search_files)
+- Looking for a specific string literal (use ripgrep)
+- Working with non-code files
+
+## Core Workflows
+
+### 1. Initialize a Project
+
+```bash
+# Initialize Graphify for a project directory
+cd /path/to/project
+graphify init
+
+# Index the codebase (builds the knowledge graph)
+graphify index
+
+# Index with specific languages only
+graphify index --lang python,typescript
+```
+
+### 2. Query the Knowledge Graph
+
+```bash
+# Natural language query
+graphify query "What services call database methods?"
+
+# Find all dependencies of a module
+graphify deps UserService
+
+# Find all callers of a function
+graphify callers main
+graphify callers "DatabaseService.save"
+
+# Trace a call chain
+graphify trace "APIHandler.process_request" --depth 5
+
+# Find shortest path between two nodes
+graphify path "APIHandler" "DatabaseConnection"
+```
+
+### 3. Structured Output (for LLM consumption)
+
+```bash
+# JSON output — pipe directly to LLM
+graphify query "What depends on the auth module?" --format json
+
+# Streaming JSON for large results
+graphify deps LargeModule --format json --stream
+```
+
+### 4. Git Integration
+
+```bash
+# Enable auto-refresh on git operations
+graphify hooks install
+
+# Manual refresh after changes
+graphify refresh
+```
+
+## Hermes Agent Integration
+
+### Code Understanding Before Changes
+
+Before modifying code, use Graphify to understand the full picture:
+
+```bash
+# 1. Understand what you're touching
+graphify deps "module_name" --format json
+
+# 2. Find all callers (who depends on this?)
+graphify callers "function_name" --format json
+
+# 3. Check refactoring safety
+graphify path "entry_point" "target_module" --format json
+```
+
+### Inject into System Prompt
+
+For complex tasks, inject the codebase structure into the agent's context:
+
+```bash
+# Get a structural overview
+graphify query "Give me a high-level overview of the project structure" --format json
+```
+
+Then include the JSON output in the system message or user message to ground the agent's understanding.
+
+### With Test-Driven Development
+
+Use Graphify to understand test impact:
+
+```bash
+# What code does this test exercise?
+graphify callers "TestAuth.test_login" --depth 3 --format json
+
+# What tests cover this function?
+graphify callers "AuthService.authenticate" --format json | grep test
+```
+
+### With Systematic Debugging
+
+Trace bugs through the call chain:
+
+```bash
+# Where does this error originate?
+graphify trace "APIHandler.handle_request" --format json
+
+# What calls this failing function?
+graphify callers "PaymentService.charge" --depth 5
+```
+
+## Tips
+
+- **Run `graphify index` after pulling changes** — the graph goes stale otherwise
+- **Use `--format json` for agent integration** — structured output is easier to consume
+- **Combine with `search_files`** — use Graphify for structure, ripgrep for content
+- **Cache is project-local** — each project has its own graph, no cross-contamination
+- **Large repos?** Index only the directories you're working in: `graphify index --path src/core`
+
+## Troubleshooting
+
+| Problem | Solution |
+|---------|----------|
+| "graphify: command not found" | `pip install graphify-cg` |
+| "No graph found" | Run `graphify init && graphify index` first |
+| Stale results | Run `graphify refresh` |
+| Language not supported | Check `graphify --help` for supported list |
+| Slow indexing | Use `--path` to limit scope |

tests/gateway/test_weak_credential_guard.py

@@ -0,0 +1,52 @@
+"""Tests for weak credential guard in gateway/config.py."""
+
+import os
+import pytest
+
+from gateway.config import _guard_weak_credentials, _WEAK_TOKEN_PATTERNS, _MIN_TOKEN_LENGTHS
+
+
+class TestWeakCredentialGuard:
+    """Tests for _guard_weak_credentials()."""
+
+    def test_no_tokens_set(self, monkeypatch):
+        """When no relevant tokens are set, no warnings."""
+        for var in _MIN_TOKEN_LENGTHS:
+            monkeypatch.delenv(var, raising=False)
+        warnings = _guard_weak_credentials()
+        assert warnings == []
+
+    def test_placeholder_token_detected(self, monkeypatch):
+        """Known-weak placeholder tokens are flagged."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "your-token-here")
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "TELEGRAM_BOT_TOKEN" in warnings[0]
+        assert "placeholder" in warnings[0].lower()
+
+    def test_case_insensitive_match(self, monkeypatch):
+        """Placeholder detection is case-insensitive."""
+        monkeypatch.setenv("DISCORD_BOT_TOKEN", "FAKE")
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "DISCORD_BOT_TOKEN" in warnings[0]
+
+    def test_short_token_detected(self, monkeypatch):
+        """Suspiciously short tokens are flagged."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "abc123")  # 6 chars, min is 30
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 1
+        assert "short" in warnings[0].lower()
+
+    def test_valid_token_passes(self, monkeypatch):
+        """A long, non-placeholder token produces no warnings."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "1234567890:ABCDEFGHIJKLMNOPQRSTUVWXYZ1234567")
+        warnings = _guard_weak_credentials()
+        assert warnings == []
+
+    def test_multiple_weak_tokens(self, monkeypatch):
+        """Multiple weak tokens each produce a warning."""
+        monkeypatch.setenv("TELEGRAM_BOT_TOKEN", "change-me")
+        monkeypatch.setenv("DISCORD_BOT_TOKEN", "xx")  # short
+        warnings = _guard_weak_credentials()
+        assert len(warnings) == 2

tests/plugins/memory/test_temporal_decay.py

@@ -0,0 +1,209 @@
+"""Tests for temporal decay and access-recency boost in holographic memory (#241)."""
+
+import math
+from datetime import datetime, timedelta, timezone
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+class TestTemporalDecay:
+    """Test _temporal_decay exponential decay formula."""
+
+    def _make_retriever(self, half_life=60):
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        return FactRetriever(store=store, temporal_decay_half_life=half_life)
+
+    def test_fresh_fact_no_decay(self):
+        """A fact updated today should have decay ≈ 1.0."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc).isoformat()
+        decay = r._temporal_decay(now)
+        assert decay > 0.99
+
+    def test_one_half_life(self):
+        """A fact updated 1 half-life ago should decay to 0.5."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=60)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.5) < 0.01
+
+    def test_two_half_lives(self):
+        """A fact updated 2 half-lives ago should decay to 0.25."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=120)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.25) < 0.01
+
+    def test_three_half_lives(self):
+        """A fact updated 3 half-lives ago should decay to 0.125."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=180)).isoformat()
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.125) < 0.01
+
+    def test_half_life_disabled(self):
+        """When half_life=0, decay should always be 1.0."""
+        r = self._make_retriever(half_life=0)
+        old = (datetime.now(timezone.utc) - timedelta(days=365)).isoformat()
+        assert r._temporal_decay(old) == 1.0
+
+    def test_none_timestamp(self):
+        """Missing timestamp should return 1.0 (no decay)."""
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay(None) == 1.0
+
+    def test_empty_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay("") == 1.0
+
+    def test_invalid_timestamp(self):
+        """Malformed timestamp should return 1.0 (fail open)."""
+        r = self._make_retriever(half_life=60)
+        assert r._temporal_decay("not-a-date") == 1.0
+
+    def test_future_timestamp(self):
+        """Future timestamp should return 1.0 (no decay for future dates)."""
+        r = self._make_retriever(half_life=60)
+        future = (datetime.now(timezone.utc) + timedelta(days=10)).isoformat()
+        assert r._temporal_decay(future) == 1.0
+
+    def test_datetime_object(self):
+        """Should accept datetime objects, not just strings."""
+        r = self._make_retriever(half_life=60)
+        old = datetime.now(timezone.utc) - timedelta(days=60)
+        decay = r._temporal_decay(old)
+        assert abs(decay - 0.5) < 0.01
+
+    def test_different_half_lives(self):
+        """30-day half-life should decay faster than 90-day."""
+        r30 = self._make_retriever(half_life=30)
+        r90 = self._make_retriever(half_life=90)
+        old = (datetime.now(timezone.utc) - timedelta(days=45)).isoformat()
+        assert r30._temporal_decay(old) < r90._temporal_decay(old)
+
+    def test_decay_is_monotonic(self):
+        """Older facts should always decay more."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc)
+        d1 = r._temporal_decay((now - timedelta(days=10)).isoformat())
+        d2 = r._temporal_decay((now - timedelta(days=30)).isoformat())
+        d3 = r._temporal_decay((now - timedelta(days=60)).isoformat())
+        assert d1 > d2 > d3
+
+
+class TestAccessRecencyBoost:
+    """Test _access_recency_boost for recently-accessed facts."""
+
+    def _make_retriever(self, half_life=60):
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        return FactRetriever(store=store, temporal_decay_half_life=half_life)
+
+    def test_just_accessed_max_boost(self):
+        """A fact accessed just now should get maximum boost (1.5)."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc).isoformat()
+        boost = r._access_recency_boost(now)
+        assert boost > 1.45  # Near 1.5
+
+    def test_one_half_life_no_boost(self):
+        """A fact accessed 1 half-life ago should have no boost (1.0)."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=60)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert abs(boost - 1.0) < 0.01
+
+    def test_half_way_boost(self):
+        """A fact accessed 0.5 half-lives ago should get ~1.25 boost."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=30)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert abs(boost - 1.25) < 0.05
+
+    def test_beyond_one_half_life_no_boost(self):
+        """Beyond 1 half-life, boost should be 1.0."""
+        r = self._make_retriever(half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=90)).isoformat()
+        boost = r._access_recency_boost(old)
+        assert boost == 1.0
+
+    def test_disabled_no_boost(self):
+        """When half_life=0, boost should be 1.0."""
+        r = self._make_retriever(half_life=0)
+        now = datetime.now(timezone.utc).isoformat()
+        assert r._access_recency_boost(now) == 1.0
+
+    def test_none_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._access_recency_boost(None) == 1.0
+
+    def test_invalid_timestamp(self):
+        r = self._make_retriever(half_life=60)
+        assert r._access_recency_boost("bad") == 1.0
+
+    def test_boost_range(self):
+        """Boost should always be in [1.0, 1.5]."""
+        r = self._make_retriever(half_life=60)
+        now = datetime.now(timezone.utc)
+        for days in [0, 1, 15, 30, 45, 59, 60, 90, 365]:
+            ts = (now - timedelta(days=days)).isoformat()
+            boost = r._access_recency_boost(ts)
+            assert 1.0 <= boost <= 1.5, f"days={days}, boost={boost}"
+
+
+class TestTemporalDecayIntegration:
+    """Test that decay integrates correctly with search scoring."""
+
+    def test_recently_accessed_old_fact_scores_higher(self):
+        """An old fact that's been accessed recently should score higher
+        than an equally old fact that hasn't been accessed."""
+        from plugins.memory.holographic.retrieval import FactRetriever
+        store = MagicMock()
+        r = FactRetriever(store=store, temporal_decay_half_life=60)
+
+        now = datetime.now(timezone.utc)
+        old_date = (now - timedelta(days=120)).isoformat()  # 2 half-lives old
+        recent_access = (now - timedelta(days=10)).isoformat()  # accessed 10 days ago
+        old_access = (now - timedelta(days=200)).isoformat()  # accessed 200 days ago
+
+        # Old fact, recently accessed
+        decay1 = r._temporal_decay(old_date)
+        boost1 = r._access_recency_boost(recent_access)
+        effective1 = min(1.0, decay1 * boost1)
+
+        # Old fact, not recently accessed
+        decay2 = r._temporal_decay(old_date)
+        boost2 = r._access_recency_boost(old_access)
+        effective2 = min(1.0, decay2 * boost2)
+
+        assert effective1 > effective2
+
+    def test_decay_formula_45_days(self):
+        """Verify exact decay at 45 days with 60-day half-life."""
+        from plugins.memory.holographic.retrieval import FactRetriever
+        r = FactRetriever(store=MagicMock(), temporal_decay_half_life=60)
+        old = (datetime.now(timezone.utc) - timedelta(days=45)).isoformat()
+        decay = r._temporal_decay(old)
+        expected = math.pow(0.5, 45/60)
+        assert abs(decay - expected) < 0.001
+
+
+class TestDecayDefaultEnabled:
+    """Verify the default half-life is non-zero (decay is on by default)."""
+
+    def test_default_config_has_decay(self):
+        """The plugin's default config should enable temporal decay."""
+        from plugins.memory.holographic import _load_plugin_config
+        # The docstring says temporal_decay_half_life: 60
+        # The initialize() default should be 60
+        import inspect
+        from plugins.memory.holographic import HolographicMemoryProvider
+        src = inspect.getsource(HolographicMemoryProvider.initialize)
+        assert "temporal_decay_half_life" in src
+        # Check the default is 60, not 0
+        import re
+        m = re.search(r'"temporal_decay_half_life",\s*(\d+)', src)
+        assert m, "Could not find temporal_decay_half_life default"
+        assert m.group(1) == "60", f"Default is {m.group(1)}, expected 60"

tests/test_model_tools.py

@@ -137,3 +137,78 @@ class TestBackwardCompat:
     def test_tool_to_toolset_map(self):
         assert isinstance(TOOL_TO_TOOLSET_MAP, dict)
         assert len(TOOL_TO_TOOLSET_MAP) > 0
+
+
+class TestToolReturnTypeValidation:
+    """Poka-yoke: tool handlers must return JSON strings."""
+
+    def test_handler_returning_dict_is_wrapped(self, monkeypatch):
+        """A handler that returns a dict should be auto-wrapped to JSON string."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        # Register a bad handler that returns dict instead of str
+        registry.register(
+            name="__test_bad_dict",
+            toolset="test",
+            schema={"name": "__test_bad_dict", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: {"this is": "a dict not a string"},
+        )
+        result = handle_function_call("__test_bad_dict", {})
+        parsed = json.loads(result)
+        assert "output" in parsed
+        assert "_type_warning" in parsed
+        # Cleanup
+        registry._tools.pop("__test_bad_dict", None)
+
+    def test_handler_returning_none_is_wrapped(self, monkeypatch):
+        """A handler that returns None should be auto-wrapped."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_bad_none",
+            toolset="test",
+            schema={"name": "__test_bad_none", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: None,
+        )
+        result = handle_function_call("__test_bad_none", {})
+        parsed = json.loads(result)
+        assert "_type_warning" in parsed
+        registry._tools.pop("__test_bad_none", None)
+
+    def test_handler_returning_non_json_string_is_wrapped(self):
+        """A handler returning a plain string (not JSON) should be wrapped."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_bad_plain",
+            toolset="test",
+            schema={"name": "__test_bad_plain", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: "just a plain string, not json",
+        )
+        result = handle_function_call("__test_bad_plain", {})
+        parsed = json.loads(result)
+        assert "output" in parsed
+        registry._tools.pop("__test_bad_plain", None)
+
+    def test_handler_returning_valid_json_passes_through(self):
+        """A handler returning valid JSON string passes through unchanged."""
+        from tools.registry import registry
+        from model_tools import handle_function_call
+        import json
+
+        registry.register(
+            name="__test_good",
+            toolset="test",
+            schema={"name": "__test_good", "description": "test", "parameters": {"type": "object", "properties": {}}},
+            handler=lambda args, **kw: json.dumps({"status": "ok", "data": [1, 2, 3]}),
+        )
+        result = handle_function_call("__test_good", {})
+        parsed = json.loads(result)
+        assert parsed == {"status": "ok", "data": [1, 2, 3]}
+        registry._tools.pop("__test_good", None)

tests/tools/test_browser_use_provider.py

@@ -0,0 +1,246 @@
+"""Tests for Browser Use provider — anti-detect, CAPTCHA, profiles, persistence."""
+
+from unittest.mock import patch, MagicMock
+import pytest
+
+
+class TestBrowserUseProviderConfig:
+    """Test configuration resolution for Browser Use provider."""
+
+    def test_not_configured_without_key(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_USE_API_KEY", raising=False)
+        from tools.browser_providers.browser_use import BrowserUseProvider
+
+        provider = BrowserUseProvider()
+        assert provider.is_configured() is False
+
+    def test_configured_with_api_key(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_USE_API_KEY", "test-key-123")
+        from tools.browser_providers.browser_use import BrowserUseProvider
+
+        provider = BrowserUseProvider()
+        assert provider.is_configured() is True
+
+    def test_provider_name(self):
+        from tools.browser_providers.browser_use import BrowserUseProvider
+
+        assert BrowserUseProvider().provider_name() == "Browser Use"
+
+
+class TestBrowserUseFeatures:
+    """Test feature configuration and payload assembly."""
+
+    def _make_provider(self, monkeypatch, **env):
+        defaults = {
+            "BROWSER_USE_API_KEY": "test-key",
+            "BROWSER_USE_ANTI_DETECT": "true",
+            "BROWSER_USE_CAPTCHA_SOLVING": "true",
+            "BROWSER_USE_PERSISTENT_LOGIN": "false",
+            "BROWSER_USE_PROFILE_NAME": "",
+            "BROWSER_USE_PROXY_COUNTRY": "us",
+        }
+        defaults.update(env)
+        for k, v in defaults.items():
+            if v is None:
+                monkeypatch.delenv(k, raising=False)
+            else:
+                monkeypatch.setenv(k, v)
+
+        # Reimport to pick up new env values
+        import importlib
+        import tools.browser_providers.browser_use as mod
+        importlib.reload(mod)
+        return mod.BrowserUseProvider()
+
+    def test_features_all_enabled(self, monkeypatch):
+        provider = self._make_provider(
+            monkeypatch,
+            BROWSER_USE_ANTI_DETECT="true",
+            BROWSER_USE_CAPTCHA_SOLVING="true",
+            BROWSER_USE_PERSISTENT_LOGIN="true",
+            BROWSER_USE_PROFILE_NAME="my-profile",
+        )
+        features = provider._build_features()
+        assert features["anti_detect"] is True
+        assert features["captcha_solving"] is True
+        assert features["persistent_login"] is True
+        assert features["profile"] is True
+        assert features["proxy_country"] == "us"
+
+    def test_features_all_disabled(self, monkeypatch):
+        provider = self._make_provider(
+            monkeypatch,
+            BROWSER_USE_ANTI_DETECT="false",
+            BROWSER_USE_CAPTCHA_SOLVING="false",
+            BROWSER_USE_PERSISTENT_LOGIN="false",
+            BROWSER_USE_PROFILE_NAME="",
+        )
+        features = provider._build_features()
+        assert features["anti_detect"] is False
+        assert features["captcha_solving"] is False
+        assert features["persistent_login"] is False
+        assert features["profile"] is False
+
+    def test_payload_includes_anti_detect(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_ANTI_DETECT="true")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert payload.get("antiDetect") is True
+
+    def test_payload_excludes_anti_detect_when_disabled(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_ANTI_DETECT="false")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert "antiDetect" not in payload
+
+    def test_payload_includes_captcha(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_CAPTCHA_SOLVING="true")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert payload.get("captchaSolving") is True
+
+    def test_payload_includes_persistent_login(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_PERSISTENT_LOGIN="true")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert payload.get("keepCookies") is True
+
+    def test_payload_includes_profile(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_PROFILE_NAME="work-account")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert payload.get("profileName") == "work-account"
+
+    def test_payload_includes_proxy_country(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_PROXY_COUNTRY="de")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert payload.get("proxyCountryCode") == "de"
+
+    def test_managed_mode_payload(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+        payload = provider._build_create_payload(managed_mode=True)
+        assert payload["timeout"] == 5
+        assert payload["proxyCountryCode"] == "us"
+
+    def test_empty_profile_excluded(self, monkeypatch):
+        provider = self._make_provider(monkeypatch, BROWSER_USE_PROFILE_NAME="")
+        payload = provider._build_create_payload(managed_mode=False)
+        assert "profileName" not in payload
+
+
+class TestBrowserUseSessionCreation:
+    """Test session creation with mocked HTTP."""
+
+    def _make_provider(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_USE_API_KEY", "test-key")
+        import importlib
+        import tools.browser_providers.browser_use as mod
+        importlib.reload(mod)
+        return mod.BrowserUseProvider()
+
+    def test_create_session_returns_features(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.ok = True
+        mock_response.json.return_value = {
+            "id": "sess-123",
+            "cdpUrl": "wss://cdp.browser-use.com/session/123",
+        }
+        mock_response.headers = {}
+
+        with patch("tools.browser_providers.browser_use.requests.post", return_value=mock_response):
+            result = provider.create_session("task-1")
+
+        assert result["bb_session_id"] == "sess-123"
+        assert result["features"]["browser_use"] is True
+        assert result["features"]["anti_detect"] is True
+        assert result["features"]["captcha_solving"] is True
+
+    def test_create_session_sends_payload(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.ok = True
+        mock_response.json.return_value = {"id": "sess-456", "cdpUrl": "wss://cdp.example.com"}
+        mock_response.headers = {}
+
+        captured_payload = {}
+
+        def mock_post(url, **kwargs):
+            captured_payload.update(kwargs.get("json", {}))
+            return mock_response
+
+        with patch("tools.browser_providers.browser_use.requests.post", side_effect=mock_post):
+            provider.create_session("task-2")
+
+        assert captured_payload.get("antiDetect") is True
+        assert captured_payload.get("captchaSolving") is True
+
+    def test_create_session_with_profile(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_USE_PROFILE_NAME", "shared-team")
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.ok = True
+        mock_response.json.return_value = {"id": "sess-789", "cdpUrl": "wss://cdp.example.com"}
+        mock_response.headers = {}
+
+        captured_payload = {}
+
+        def mock_post(url, **kwargs):
+            captured_payload.update(kwargs.get("json", {}))
+            return mock_response
+
+        with patch("tools.browser_providers.browser_use.requests.post", side_effect=mock_post):
+            provider.create_session("task-3")
+
+        assert captured_payload.get("profileName") == "shared-team"
+
+    def test_create_session_error_raises(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.ok = False
+        mock_response.status_code = 401
+        mock_response.text = "Unauthorized"
+
+        with patch("tools.browser_providers.browser_use.requests.post", return_value=mock_response):
+            with pytest.raises(RuntimeError, match="Failed to create Browser Use session"):
+                provider.create_session("task-err")
+
+
+class TestBrowserUseSessionCleanup:
+    """Test session cleanup."""
+
+    def _make_provider(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_USE_API_KEY", "test-key")
+        import importlib
+        import tools.browser_providers.browser_use as mod
+        importlib.reload(mod)
+        return mod.BrowserUseProvider()
+
+    def test_close_session_success(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.status_code = 200
+
+        with patch("tools.browser_providers.browser_use.requests.patch", return_value=mock_response):
+            result = provider.close_session("sess-123")
+
+        assert result is True
+
+    def test_close_session_failure(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        mock_response = MagicMock()
+        mock_response.status_code = 500
+        mock_response.text = "Internal Server Error"
+
+        with patch("tools.browser_providers.browser_use.requests.patch", return_value=mock_response):
+            result = provider.close_session("sess-bad")
+
+        assert result is False
+
+    def test_emergency_cleanup_no_error_on_failure(self, monkeypatch):
+        provider = self._make_provider(monkeypatch)
+
+        with patch("tools.browser_providers.browser_use.requests.patch", side_effect=Exception("network down")):
+            # Should not raise
+            provider.emergency_cleanup("sess-xyz")

tools/browser_providers/browser_use.py

@@ -1,4 +1,8 @@
-"""Browser Use cloud browser provider."""
+"""Browser Use cloud browser provider.
+
+Enhanced with anti-detect profiles, CAPTCHA solving, persistent logins,
+and profile management per Issue #262.
+"""
 
 import logging
 import os
@@ -21,6 +25,17 @@ _DEFAULT_MANAGED_TIMEOUT_MINUTES = 5
 _DEFAULT_MANAGED_PROXY_COUNTRY_CODE = "us"
 
 
+# ---------------------------------------------------------------------------
+# Config helpers
+# ---------------------------------------------------------------------------
+
+def _env_bool(key: str, default: bool = False) -> bool:
+    val = os.environ.get(key)
+    if val is None:
+        return default
+    return val.lower() in ("1", "true", "yes", "on")
+
+
 def _get_or_create_pending_create_key(task_id: str) -> str:
     with _pending_create_keys_lock:
         existing = _pending_create_keys.get(task_id)
@@ -61,7 +76,28 @@ def _should_preserve_pending_create_key(response: requests.Response) -> bool:
 
 
 class BrowserUseProvider(CloudBrowserProvider):
-    """Browser Use (https://browser-use.com) cloud browser backend."""
+    """Browser Use (https://browser-use.com) cloud browser backend.
+
+    Supports anti-detect profiles, CAPTCHA solving, persistent logins,
+    and named profile management.  Configuration via env vars:
+
+    - ``BROWSER_USE_API_KEY``          — direct API key
+    - ``BROWSER_USE_ANTI_DETECT``      — enable anti-detect fingerprinting (default: true)
+    - ``BROWSER_USE_CAPTCHA_SOLVING``  — enable CAPTCHA auto-solving (default: true)
+    - ``BROWSER_USE_PERSISTENT_LOGIN`` — persist cookies across sessions (default: false)
+    - ``BROWSER_USE_PROFILE_NAME``     — named profile for cookie jar isolation
+    - ``BROWSER_USE_PROXY_COUNTRY``    — proxy country code override (default: us)
+    """
+
+    # Feature config snapshot — read once at import so runtime env mutation
+    # cannot silently change an in-flight session's capabilities.
+    _cfg_anti_detect: bool = _env_bool("BROWSER_USE_ANTI_DETECT", True)
+    _cfg_captcha: bool = _env_bool("BROWSER_USE_CAPTCHA_SOLVING", True)
+    _cfg_persistent: bool = _env_bool("BROWSER_USE_PERSISTENT_LOGIN", False)
+    _cfg_profile: str = os.environ.get("BROWSER_USE_PROFILE_NAME", "")
+    _cfg_proxy_country: str = os.environ.get(
+        "BROWSER_USE_PROXY_COUNTRY", _DEFAULT_MANAGED_PROXY_COUNTRY_CODE
+    )
 
     def provider_name(self) -> str:
         return "Browser Use"
@@ -106,6 +142,56 @@ class BrowserUseProvider(CloudBrowserProvider):
             raise ValueError(message)
         return config
 
+    # ------------------------------------------------------------------
+    # Feature / payload assembly
+    # ------------------------------------------------------------------
+
+    def _build_features(self) -> Dict[str, Any]:
+        """Return a dict describing which features are active."""
+        return {
+            "browser_use": True,
+            "anti_detect": self._cfg_anti_detect,
+            "captcha_solving": self._cfg_captcha,
+            "persistent_login": self._cfg_persistent,
+            "profile": bool(self._cfg_profile),
+            "proxy_country": self._cfg_proxy_country,
+        }
+
+    def _build_create_payload(self, managed_mode: bool) -> Dict[str, Any]:
+        """Build the session creation payload with all configured features."""
+        payload: Dict[str, Any] = {}
+
+        if managed_mode:
+            payload["timeout"] = _DEFAULT_MANAGED_TIMEOUT_MINUTES
+            payload["proxyCountryCode"] = self._cfg_proxy_country
+        elif self._cfg_proxy_country:
+            payload["proxyCountryCode"] = self._cfg_proxy_country
+
+        # Anti-detect fingerprinting — Browser Use v3 uses
+        # the antiDetect field to enable browser fingerprint spoofing
+        # so sites see a real human browser, not an automation tool.
+        if self._cfg_anti_detect:
+            payload["antiDetect"] = True
+
+        # CAPTCHA solving — Browser Use handles reCAPTCHA, hCaptcha,
+        # Cloudflare Turnstile, and custom CAPTCHAs automatically.
+        if self._cfg_captcha:
+            payload["captchaSolving"] = True
+
+        # Persistent login — preserves cookies across sessions when
+        # a named profile is provided, enabling sites like Gmail, Twitter,
+        # etc. to stay logged in between agent runs.
+        if self._cfg_persistent:
+            payload["keepCookies"] = True
+
+        # Named profile — isolates cookie jar and browser state under
+        # a specific profile name.  Multiple agents can share a profile
+        # or use isolated profiles for different accounts.
+        if self._cfg_profile:
+            payload["profileName"] = self._cfg_profile
+
+        return payload
+
     # ------------------------------------------------------------------
     # Session lifecycle
     # ------------------------------------------------------------------
@@ -125,17 +211,7 @@ class BrowserUseProvider(CloudBrowserProvider):
         if managed_mode:
             headers["X-Idempotency-Key"] = _get_or_create_pending_create_key(task_id)
 
-        # Keep gateway-backed sessions short so billing authorization does not
-        # default to a long Browser-Use timeout when Hermes only needs a task-
-        # scoped ephemeral browser.
-        payload = (
-            {
-                "timeout": _DEFAULT_MANAGED_TIMEOUT_MINUTES,
-                "proxyCountryCode": _DEFAULT_MANAGED_PROXY_COUNTRY_CODE,
-            }
-            if managed_mode
-            else {}
-        )
+        payload = self._build_create_payload(managed_mode)
 
         response = requests.post(
             f"{config['base_url']}/browsers",
@@ -158,7 +234,14 @@ class BrowserUseProvider(CloudBrowserProvider):
         session_name = f"hermes_{task_id}_{uuid.uuid4().hex[:8]}"
         external_call_id = response.headers.get("x-external-call-id") if managed_mode else None
 
-        logger.info("Created Browser Use session %s", session_name)
+        logger.info(
+            "Created Browser Use session %s [anti_detect=%s captcha=%s persistent=%s profile=%s]",
+            session_name,
+            self._cfg_anti_detect,
+            self._cfg_captcha,
+            self._cfg_persistent,
+            self._cfg_profile or "(none)",
+        )
 
         cdp_url = session_data.get("cdpUrl") or session_data.get("connectUrl") or ""
 
@@ -166,7 +249,7 @@ class BrowserUseProvider(CloudBrowserProvider):
             "session_name": session_name,
             "bb_session_id": session_data["id"],
             "cdp_url": cdp_url,
-            "features": {"browser_use": True},
+            "features": self._build_features(),
             "external_call_id": external_call_id,
         }